Fast noisy image quality assessment based on free-energy principle

In this work, we propose a fast noisy image quality assessment approach under the theory of free-energy principle. The free-energy principle accounts for that the brain tries to predict the input image with an internal generative model. While there exists a discrepancy between the image and its model-predicted version and this discrepancy is believed to be closely related to the perceptual quality of the image. Accordingly, we devise the method for evaluating the quality of the noisy image, in which the internal generative model is first generalized from AR simulation and then instantiated with simple but effective median function. The quality-related discrepancy denoted by the entropy of the predicted residuals is defined to measure the whole quality of the image. Experimental results on LIVE, TID2013 and CSIQ databases demonstrate that the proposed method earns comparable prediction performance to the specialized noise level estimation methods while greatly reduces the running time.